Re: Issues with exceptions and overloading opcodes

On Tue, 20 Oct 2015, Bob Weinand wrote: > > Am 20.10.2015 um 18:08 schrieb Derick Rethans <derick@php.net>: > > > > On Tue, 20 Oct 2015, Bob Weinand wrote: > > > >>> Am 20.10.2015 um 17:11 schrieb Xinchen Hui <laruence@php.net>: > >>> > >>> On Tue, Oct 20, 2015 at 10:45 PM, Derick Rethans <derick@php.net > >>> <mailto:derick@php.net>> wrote: > >>> > >>>> When hacking on Xdebug for PHP 7, I ran into some issues where > >>>> PHP would just spin around. This happens when there is an > >>>> exception, and I use overloaded opcodes. I wrote a small example > >>>> extension at > >>>> > >>>> https://github.com/derickr/php-minimal-opcode-overloading-example > >>>> with a test case ( > >>>> > >>>> https://github.com/derickr/php-minimal-opcode-overloading-example/blob/master/tests/test.php > >>>> > >>>> ) that shows that something is looping in executing opcodes, as > >>>> the HANDLE_EXCEPTION iirc doesn't advance to the next opline. I > >>>> believe this is a recent enough change, as it only started > >>>> happening after I upgraded from about RC1 to latest master. > >>>> > >>>> Would you care to have a look? > >>>> > >>> This is introduced by a fix made by Bob, > >>> https://github.com/php/php-src/commit/808f62bb > >>> > >>> <https://github.com/php/php-src/commit/808f62bb> > >>> > >>> Bob, I am going to revert this for now, since you didn't include a > >>> test script to show where the problem was, I am not sure why you > >>> made this? > >> > >> I didn't add any test case as we don't have any APIs in php-src > >> using that user_opcode handler. I discussed the change back then > >> with Dmitry and committed with his review. > >> > >> It concretely broke uopz which might throw an exception itself. And > >> in turn we realized that for integrity we need to ensure that > >> opline == EX(opline) else we'll end up with different behavior in > >> global register builds and normal builds. > >> > >> In general, you should not overload ZEND_HANDLE_EXCEPTION (if you > >> need to, have a special handling for it, but never directly > >> dispatch to it). It is quite special being the only opcode actually > >> allowing EG(exception) to be non-NULL at the start of it. > >> > >> You are doing an #ifdef ZTS to exempt it… why don't you always > >> exempt it? > > > > Even if I do (like I've just pushed into the extension), the test > > still fails because the code loops. So you did break something. > > Ah, ZEND_CATCH indeed also allows EG(exception) to be non-null… > > … maybe I should write > > if (EG(exception) > && opline->opcode != ZEND_HANDLE_EXCEPTION && opline->opcode != > ZEND_CATCH > && ret != ZEND_USER_OPCODE_DISPATCH && ret != ZEND_USER_OPCODE_RETURN > && ret != ZEND_USER_OPCODE_ENTER && ret != ZEND_USER_OPCODE_LEAVE) { > HANDLE_EXCEPTION(); > } > > to allow dispatching for these two opcodes too? I don't know. Dmitry, Xinchen? I'd like to get this resolved soon.. cheers, Derick